A reproducible endothelial-state contrast with limited bulk prognostic robustness in hepatocellular carcinoma

Endothelial-cell (EC) states may influence the hepatocellular carcinoma (HCC) microenvironment, but EC-derived bulk-tissue scores can also capture tumour purity and cellular admixture. We re-derived and stress-tested a previously proposed vascular immune accessibility score. GSE149614 was analysed with sample-aware doublet detection, ambient-RNA correction, pan-cell-type specificity testing, and patient-level pseudobulk contrasts. A deterministic algorithm selected two 15-gene modules. Independent single-cell validation used GSE151530, with all score genes excluded from unsupervised tumour endothelial-cell (TEC) clustering. Bulk sensitivity analyses used TCGA-LIHC, GSE14520, and ICGC LIRI-JP; TCGA models incorporated ABSOLUTE purity. Spatial analyses used two unique GSE245908 sections. Only 9/25 original immune-adhesive genes and 6/25 barrier genes passed EC-specificity thresholds. GSE151530 included 50,023 HCC cells from 25 patients and 32 specimens, including 2,299 quality-controlled TECs. At the patient level, both modules were EC enriched (median TEC-minus-non-TEC differences 1.438 and 3.222; paired Wilcoxon P = 2.21 × 10^-5 and 1.94 × 10^-5). The modules mapped onto 12 independently derived TEC clusters. The revised score correlated modestly with TCGA ABSOLUTE purity (rho=-0.137; P = 0.010). Clinical Cox models were non-significant in TCGA-LIHC, GSE14520, and ICGC LIRI-JP, and cohort directions were inconsistent. Spatial results replicated in only one of two sections. The revised score was not evaluable in the 24-sample NanoString immunotherapy cohort because only 3/30 genes were present. The reproducible result is an LSEC-like-to-angiogenic EC transcriptomic contrast. The data do not support direct immune-accessibility, directionally consistent prognostic, or immunotherapy-selection claims.